Improving Direct Image Regression for Blood Cell Enumeration with a Fine-Tuned Backbone

Complete blood count (CBC) examination provides an important insight for diagnosis or disease treatment. Currently, CBC examination requires complex and expensive devices that limit their deployment in remote area. The development of computer vision based method offers simplification to the process. However, its implementation is limited to the availability of large size labelled dataset. This research aims to develop a direct image regressor that is able to regress directly from image. There are two stages in estimation process. First, the backbone is trained using large dataset available for blood cell classification problem. Then the trained backbone is plugged into the final model by adding a fully connected neural network that acts as regressor. The whole model is then trained using limited whole blood cell count dataset. The evaluation process shows that training the backbone using large size related dataset improve the performance by 50%. This study can be used to create a low-cost blood component evaluation tool, particularly in rural areas where access to advanced laboratory equipment is limited.